A Magnetism-Aware Crystal Graph Neural Network with Explicit Magnetic-Sublattice and Exchange-Geometry Representations

Abstract

Magnetic order in crystals is governed by moment-carrying sublattices and ligand-mediated exchange pathways, yet standard crystal graph neural networks do not explicitly distinguish magnetic sublattices and encode bonds primarily through pair distances. We propose mCGCNN, a magnetism-aware crystal graph neural network that augments the full structural graph with a dedicated magnetic subgraph. The magnetic stream performs angle-aware message passing between magnetic centers using metal--ligand--metal exchange-path descriptors motivated by Goodenough--Kanamori--Anderson physics, while layer-wise cross-coupling transfers structural and ligand-environment information from the full crystal graph. Separate pooling of the magnetic sublattice preserves magnetism-specific information alongside the global structural representation. We evaluate mCGCNN for total magnetic moment prediction on 49,462 crystalline materials with ferromagnetic spin configurations from the Materials Project and find that the proposed model improves agreement with DFT magnetic moments over established crystal-graph approaches without requiring a large increase in model complexity. Material-level analysis shows that the largest gains are concentrated predominantly in complex transition-metal and oxygen-containing magnetic compounds. Our results show that explicitly incorporating magnetic sublattices and exchange geometry provides a physically motivated route to improved crystal-graph representations for magnetic materials.